Fidelity-Controlled Robustness Enhancement of Blind Watermarking Schemes Using Evolutionary Computational Techniques

Designing optimal watermarking schemes is inherently an interesting and difficult problem since the three most important performance requirements of digital watermarking – fidelity, robustness and watermark capacity – are conflicting with each other. Nowadays, most watermarking schemes hide the watermark information in a heuristic manner, that is, watermarks are often embedded according to predefined rules and empirical parameter settings. Therefore, the performance of digital watermarking can only be passively decided and evaluated, rather than being actively adopted as additional clues helpful to achieve better performance in embedding modules. In this paper, watermark embedding is simulated as an optimization procedure in which optimal embedded results are obtained by using important evolutionary computation techniques – the genetic algorithms. Under the condition that fixed amount of watermark bits are hidden, in this work, the minimal fidelity requirement of embedded content can be specified by users in advance and guaranteed throughout the embedding procedure. Furthermore, concrete measures of the robustness against certain attacks are treated as the objective functions that guide the optimizing procedure. In other words, a blind watermarking scheme with application-specific data capacity, guaranteed fidelity, and theoretically optimal robustness against certain types of attacks is proposed. Experimental results clearly show that the proposed scheme possesses great performance improvements over the original one. More importantly, the proposed enhancing approach has many desired architectural characteristics, such as blind detection, asymmetric embedding/detection overheads, as well as embedding and detection in different domains. I. Problem Formulation: Optimal Watermarking In the past decade, various watermarking schemes have been proposed, and many important advances in the field of digital watermarking have been made. A comprehensive introduction to current watermarking technologies and theoretic foundations can be found in [1]. However, designing optimal watermarking schemes is still an open problem to which satisfying solutions are not yet found [2]. The 2 Chun-Hsiang Huang, Chih-Hao Shen and Ja-Ling Wu difficulties one might face while designing optimal watermarking schemes is a natural result due to the three conflicting requirements of most watermarking systems: fidelity of embedded content, robustness of the hidden information against common processes or malicious attacks, and the data capacity of hidden information. Among these requirements, capacity is often decided in advance according to the purpose of watermarking application. Therefore, as long as the predefined amount of embedded data is large enough to carry necessary information, such as identifications of content authors for owner proving or usage rules that shall be parsed by DRM-enabled consumer-electronic devices, data capacity can be simply regarded as a fixed parameter. However, even under this assumption, obtaining a reasonable trade-off between fidelity and robustness is still not as simple as one might think. In existing watermarking schemes, perceptually acceptable embedded outcomes are often produced according to predefined embedding rules based on complicated perceptual models or assumptions, and then the robustness of that scheme is empirically experimented and evaluated by performing various attacks on the embedded media. In order to clearly illustrate the relationship between traditional watermarking schemes and their performance indexes, the performance-space view of digital watermarking schemes shall be introduced and examined in the beginning. As shown in Fig. 1, any embedded result created using a certain watermarking scheme can be expressed by a point located somewhere within the space spanned by two axes representing fidelity and robustness, respectively. Roughly speaking, it is generally agreed that the higher the required fidelity is, the lower the robustness of hidden signals against attacks will be. In fact, the region that possible embedded outcomes may locate within is consequently determined after the watermarking scheme and the original/watermark pairs are chosen. Although potentially better performance of the chosen watermarking scheme may exist, traditional watermarking schemes lack the ability to exploit better embedded outcomes. II. Genetic-Algorithms and Watermarking To solve the watermarking performance optimization problem within reasonable computation time, the genetic-algorithms (GAs) based optimization techniques are used in this paper. GAs are important optimization techniques belonging to the area of evolutionary computation [3]. During GA-based optimization processes, solutions to the problem can be evaluated according to objective function values representing the degree of fitness. A population of candidate solutions will be initialized as finitelength strings the so-called chromosomes over finite alphabet. Different from conventional single-point search methods, GAs work with a population of points in a parameter space simultaneously. In practical GA-based optimizations, three GA operators reproduction, crossover, and mutation are often applied to the chromosomes repeatedly. After iteratively adopting these GA operations, nearoptimal solutions of desired parameters for the original problem can be obtained. A detail explanation of GAs can be found in [4]. Fidelity-Controlled Robustness Enhancement of Blind Watermarking Schemes Using Evolutionary Computational Techniques 3 The authors first introduced the idea of GA-based watermarking enhancement in [5]. In that paper, the performance of a DCT-domain watermarking scheme, being proposed in [6], is enhanced by applying GA operators, and the best watermark embedding positions for each DCT block can be found. [5] is the earliest publication that connects GAs together with digital watermarking. However, this enhancing scheme was criticized due to several factors that seriously limit its usage. For blind watermarking schemes, the optimal embedding positions must be delivered to the watermark detector as secret keys since the decoder cannot figure out the optimized embedding positions without the original content. However, this is usually not feasible in many important real-world applications, thus seriously limiting its applicability. As for non-blind watermarking schemes, the optimization procedure will be a time-consuming process in both embedding and detection sides. Although the authors proposed a lightweight genetic search algorithm in [7] to shorten the computation time, large amounts of computation overheads in both watermark embedding and detection are still unacceptable. In [8], a similar DCT-domain watermarking scheme is proposed, where optimization for robustness is considered. Furthermore, a GA-based spatial-domain watermarking algorithm is also proposed in [8]. However, the usage of the latter scheme is limited due to the weak robustness of its spatial domain embedding nature, and both schemes suffer from the aforementioned secret key delivery problem. In this paper, a watermarking-performance enhancement architecture possessing theoretically optimal robustness against certain types of attacks, guaranteed fidelity and application-specific data capacity will be proposed. The proposed architecture is inherently suitable for blind watermarking, and the asymmetric embedding and detection design can effectively reduce the penalty of long computation time. Furthermore, the proposed watermarking scheme has the desirable characteristic that embedding and detection can be performed in different domain, thus both direct control of fidelity in spatial domain while embedding and strong robustness against attacks in frequency domain while detecting can be realized in a single framework. The paper is organized as follows. Section III describes the proposed watermarking enhancement scheme and related implementation details. Experimental results, including the performance comparisons against original watermarking scheme under the assumption of equal data capacity, will be listed and explained in Section IV. Section V gives a brief discussion about the pros and cons of the proposed performance enhancing architecture, and section VI concludes this write-up. III. The Proposed Architecture and Implementation Details The proposed performance enhancement architecture is general and can be used to enhance the performance of various existing blind watermarking schemes. However, in order to save the implementation time and consider that frequency-domain watermarking schemes are well known for their better robustness and fidelity, a blind version of the block-DCT based image watermarking scheme originally introduced by 4 Chun-Hsiang Huang, Chih-Hao Shen and Ja-Ling Wu [6] is used to evaluate the power of this enhancing approach. [9] illustrates the details about turning the originally non-blind watermarking approach into a blind one. In this scheme, embedded watermarks are meaningful binary patterns and are randomly permuted before embedding. Watermark bits are embedded into predefined positions within the middle-band coefficients of each DCT block. The polarity between AC coefficients at selected positions and scaled DC coefficient of each block, after considering the effects of JPEG compression, is adjusted to insert the watermarking bits. Due to the blind nature of the adopted scheme in [9], the original perceptual model introduced by [6] cannot be applied directly, thus the degree of coefficient adjustment is empirically and uniformly determined. The definition of polarity for the m-th watermarking bit that will be embedded into one coefficient block is given by: